Retrying an agent step is not the same as safely resuming a 14-step workflow

msignificantdigit · reddit · 2026-07-28

Retrying a failed agent step is not the same as resuming a durable workflow

A builder working with durable execution for agents argues that “retry” and “resume” are often conflated, but they describe different reliability behaviors in production.

Using a 14-step LangGraph workflow as an example, the post shows why restarting from step one can duplicate side effects such as emails, CRM updates, or infrastructure actions. Retrying only the failed call is safer, but it still leaves open whether earlier steps completed, whether outputs were durably recorded, whether the model or prompt has changed, and whether a prior human approval still applies.

The author separates three layers:

Agent decision state — what the model saw and which branch it took.
Durable orchestration state — which steps completed and where execution can safely continue.
Idempotency — whether a tool or side effect can happen twice without damage.

The main conclusion is that checkpoints help with recovery, but they do not make external actions idempotent. Tool-level idempotency remains necessary even when the runtime tracks progress.

Original post →

Retrying an agent step is not the same as safely resuming a 14-step workflow

More from coding & agent